According to MIT News, researchers at the university’s Laboratory for Information and Decision Systems (LIDS) have devised a method to embed physical constraints into machine learning models of chaotic systems, ensuring their predictions remain within realistic bounds. The work, published in the Proceedings of the National Academy of Sciences (PNAS), addresses a long-standing problem: AI models that simulate complex, nonlinear processes such as weather patterns, climate dynamics, and turbulent flows often drift into physically impossible predictions over time.
The team, comprising MIT graduate student Sunbochen Tang and professors Navid Azizan and Themistoklis Sapsis, introduced a technique that enforces energy dissipation as a hard constraint. In chaotic systems, energy dissipation naturally keeps behavior bounded—for instance, a weather system cannot produce infinite temperatures. The new approach allows the model to learn an “energy-like function” and a boundary within which chaotic behavior is confined. A projection layer then adjusts predictions to ensure that outside this boundary, the energy-like quantity decreases, drawing the simulation back toward the bounded region.
“Chaos does not mean anything can happen. These systems often remain bounded, and their models should, too,” Azizan told MIT News. He explained that a model can match data over a short period and still fail when run for a long time, as small errors accumulate until predictions grow without limit—something real systems never do.
The key innovation, Azizan said, is “hard constraining versus soft constraining.” Previous attempts to prevent unrealistic results typically used penalty functions during training, which offer no guarantee that the model will respect physical limits when faced with new data. “Our approach is based on the concept of hard-constrained neural networks,” Azizan said, “so that what comes out is always physically consistent, even if it’s a completely new case the model hasn’t seen.”
Tang noted that simply capping predictions at arbitrary values—for example, setting a maximum temperature of 100 degrees—would keep outputs bounded but would prevent the model from learning complex features directly from data. The energy-based constraint, by contrast, lets the model capture the system’s true dynamics while guaranteeing boundedness.
The team tested the method on a range of physical systems, from simple textbook cases to highly complex multidimensional systems. According to MIT News, the new approach closely matched true outcomes, while conventional methods soon produced impossible predictions.
Sapsis said the work “paves the way for applying the most advanced machine-learning architectures for prediction and control in a plethora of real-world problems, even those characterized by intrinsic instabilities.” The method is broadly applicable to any chaotic system and is designed to work with various neural network architectures. The team has released open-source code alongside the paper.
A follow-up study by the group, presented at the Learning for Dynamics and Control conference, won the conference’s Best Paper Award.
For university students studying fields from climate science to aerospace engineering, the advance could mean more reliable AI tools for long-term forecasting and system control. “For high-stakes applications, you need guarantees about the properties that must always hold,” Azizan said. “In this work, we established one such guarantee: boundedness. Our broader goal is to make principles part of how AI models are built, not just how they are tested.”